Why does a single market outage lead to a market failure?

Carole Comerton-Forde
School of Banking and Financ, UNSW

Zhuo (Joe) Zhong
Department of Finance, University of Melbourne

When Euronext suffered an outage for a little under 3 hours on 19 October 2020, the average euro volume traded over the day in Euronext stocks fell by over 40% compared to the average daily in the previous week. During the outage itself trading virtually stopped. Given multiple trading venues compete with Euronext for the provision of trading services, why did trading activity decline so dramatically when Euronext went down? Why did traders not simply route their orders to one of the many other trading venues available to them. And what can be done to ensure future outages do not negatively impact trading activity?

In this short paper we aim to provide answers to these questions. We do this by examining the trading and market quality characteristics of Euronext stocks on the outage day, and the days immediately before and after this event. We also compare these characteristics to a set of control stocks in other European markets that did not suffer an outage. We focus on trading activity, bid ask spreads, volatility and price discovery.

Our evidence shows that in the absence of a reference price from Euronext, quoting activity on the MTFs decreases and spreads widen significantly. The primary market is critical to pricing and accounts for about 60% of price discovery in Euronext stocks on a typical day. When Euronext is not there to provide a reference price for investors and market markets, trading essentially stops. Algorithms and other trading technology need the primary market reference prices to operate. If competing venues are to successfully provide redundancy in the event of market outages, reforms are needed to ensure that an appropriate reference price can be established and that the market’s price discovery mechanisms continue to function.

Our approach

Our analysis examines trading activity, bid ask spreads, volatility and price discovery. We provide definitions of how we measure these variables in an appendix (see A1).

Relying on the old adage "a picture tells a thousand words" ––– we tell the story of the outage mostly using figures. However, we also undertake formal statistical tests using a difference-in-differences approach. A difference-in-differences approach is an econometric technique used in social sciences to try to mimic the randomised controlled trials used in the hard sciences. This approach exploits "natural experiments" where some observations (treatment) are affected by the experiment and other observations (controls) are not. Our natural experiment is the Euronext outage. The treatment impacts only Euronext stocks. And our control stocks are those that have their primary listing on another European market. The difference-in-difference means that we compare treated and control stocks on the outage day between the halted and non-halted periods. The difference in the treatment and control stocks during the halted period on the outage day, is compared to the difference in the treatment and control stocks during the non-halted period. This approach helps to identify the causal impact of the treatment and mitigate any extraneous factors. We provide further details of this, for the interested reader, in the appendix (see A2).

Sample construction and data

Our analysis covers the week prior to and following the outage day, covering the period 12 October to 23 October. Our statistical tests use all of these days, but when presenting figures, we focus on the day before and after the outage, and the outage day itself.

We examine stocks included in the STOXX Total Market Index (TMI) that trade on the primary market and the three largest Multilateral Trading Facilities (MTF): Turquoise, Chi-X and BATS, each day of our sample. For simplicity we do not consider the many other venues and mechanisms where trading can be done in Europe. We obtain trade and quote data for these trading venues from Refinitiv’s Datascope.

Our sample includes 567 stocks: 109 of which have their primary listing on a Euronext market and 458 listed on non-Euronext markets. We refer to the non-Euronext sample as the control group. In the appendix, we provide a list of the stocks included in the sample, and their primary listing (see A4).

Our choice to only examine stocks that trade in all four venues, biases our sample toward the most fragmented stocks in the market, which are larger, more active stocks. We make this choice because some of the variables we examine need frequent observations in each venue to be correctly estimated at short horizons. But this choice means that our analysis underestimates the impact of the outage for investors. The less liquid and less fragmented stocks not examined will be more adversely affected by the outage as the primary market is even more important for these stocks.

How did the outage impact on trading volume?

We begin by examining daily euro trading volume for the primary market, Turquoise, Chi-X and BATS. Figure 1 shows the daily trading volume for treatment and control stocks. Trading in Euronext stocks fell by 42% from an average daily value of €38.6m to €22.3m. Figure 1 also shows that Euronext stocks were not the only stocks to that suffer because of the outage. Activity in the control stocks also fell by around 23% from €23.1m to €17.7m. On average the control stocks trade less than the Euronext stocks. The decline in control stock volume on the outage day suggests that there may be a spill-over effect perhaps because short term trading portfolio trading strategies such as index arbitrage and pairs trading that involve Euronext stocks are not possible when Euronext is down.

Next, we consider the intraday pattern in trading in the Euronext and control stocks on the outage day, and the day before and after the outage. We examine trading between 9:00 and 17:30 Central European Time (CET).[[1]](#myfootnote1) We consider only continuous trading and exclude the opening/closing auction. Figure 2 shows that on a typical day, trading activity exhibits a flat U-shape with elevated activity around the open and close of continuous trading.[[2]](#myfootnote2) Figure 2 also shows that the Euronext accounts for around 73.9% of trading on a typical day, followed by 11%, 9.9% and 5.2% on Chi-X, BATS and Turquoise respectively. The middle pane of the top panel shows that when Euronext went down, trading also largely evaporated on the MTFs. For Euronext stocks, the average volume per minute during the outage was only €34995 compared to an average of €68881 per minute on the day before (and €55690 on the day after). [ Also please quantify the decline per minute in the control stocks and if relevant add the stats]

What happened to prices?

Given there is limited trading activity, what happened to the volatility of prices? Figure 3 reports the intraday volatility in one minute increments for Euronext and control stocks on the outage day, and the day before and after the outage. On a typical day, volatility is highest at the open and near the close of continuous trading and is also elevated at the time the US markets open. Volatility on the primary market is typically lower than on the MTFs, and volatility is on average higher in the control stocks than Euronext stocks.

The middle pain of the top panel shows that volatility increases dramatically during the market outage. Volatility which is normally in the range of 0.01% to 0.07% increases to 0.1% to 7%. There are also numerous gaps in the volatility estimates due to periods without observations. It is noteworthy that volatility is also elevated in the control stocks on the outage day.

What about bid ask spreads?

Given volume is down, what is happening to liquidity provision measured by bid ask spreads. We examine bid ask spreads measured at one-minute increments across the trading day on each of the four trading venues we consider. The top panel of Figure 4 reports the bid ask spread measured in basis points for each of our four trading venues for the day of the outage and the day before and after. The bottom panel reports the same information for the control stocks.

As expected, spreads show a distinctive reverse J-shape across the trading day. Spreads are wide in the morning due to overnight uncertainty and widen again as the end of trading approaches. On average the primary markets offer significantly tighter spreads than the MTFs. This is true both for Euronext and control stocks. The middle pane of the top panel of Figure 4 shows that on the outage day, spreads blow out when Euronext is down. The average spread rises by about 700% to 1,070%.
Table 2a provide summary statistics for the Euronext and control stocks on the outage and non-outage days. On a typical day in our sample, the average (median) spread for Euronext stocks on Euronext is 22 (16) bps compared to 60 (47) bps on BATS, 68 (34) bps on Chi-X and 79 (50) bps on Turquoise. For the control stocks the spreads are wider on average, but the primary markets also offers tighter spreads than the MTFs, with average (median) spreads of 31 (20) bps. The ordering of the MTFs differs with an average (median) spread of 88 (58) bps on Turquoise, 94 (60) bps on BATS and 97 (53) bps on Chi-X. It is worth highlighting that on the outage day not all stocks were quoted on the MTFs. Only 69, 43 and 34 stocks were quoted on Chi-X, BATS and Turquoise respectively, whereas on non-outage days they are quoting 107-108 stocks of the 109 in the sample. In contrast, spreads in the control stocks remain relatively unchanged, with a small decrease in the average and median values.

Table 2b reports the results for our difference-in-difference analysis. The coefficient for the "Treatment:Halt" captures the change in the difference between the treatment and control group when trading is halted versus not halted. This shows that after controlling for other factors, the spreads on the MTFs increased by between 450 and 770 bps due to the outage. This provides compelling evidence that when the primary market is down, traders and are less willing to post competitive prices on MTFs, despite them being available for trading.[[3]](#myfootnote3)

Does this mean the MTFs do not provide a meaningful contribution to the spread? An analysis of the frequency with which the MTFs offer a better price than the primary suggests the answer to that question is no.

Table 3 reports the percentage of time the MTFs are at either the best bid and offer price (BBO) when the primary market is not at these prices. Columns 3 and 4 report this percentage for all MTFs together, and the remaining columns report it for each MTF separately. For Euronext stocks, the MTFs are at the BBO without the primary market offering the same price approximately 27% of the trading day. For the control stocks the MTFs are at the BBO without the primary market slightly less frequently. Chi-X is more likely to be at the BBO without the primary market more often than BATS and Turquoise.

These results show that on a normal trading day, the three main MTFs provide economically meaningful improvements in spreads. Why do they not continue to do this when Euronext goes down? We suspect it is due to the importance of the primary market as a reference price. The primary market is likely to be critical to price discovery, and in its absence, traders are less willing to submit orders. So, for our final analysis we turn our attention to identifying where price discovery occurs on a typical trading day.

Where does price discovery take place?

Twenty five years ago, Joel Hasbrouck, a market structure academic, developed an econometric technique to identify where price discovery takes place. This approach has been the main tool used in the academic literature to identify the relative contribution of different trading venues to price discovery. We again leave a detailed discussion of the econometrics to the appendix, but the principle of the technique is to assign the proportion of the efficient price innovation variance that is provided by each trading venue.

We report the information share or the contribution of the primary venue to price discovery in Table 4 on the day before and after the outage. This shows that on in average (median) stock on Euronext, Euronext accounts for between 60% (56%) of price discovery, with the MTFs accounting for the remainder. Other primary markets account for marginally higher fractions of price discovery.

The dominance of the primary market in the price discovery process explains why traders rely on it as the reference price. It also offers a potential justification of why many trading algorithms and other systems rely on the primary market in order to operate. However, the consequence of this means that when the primary market suffers an outage, trading will essentially stop.

How can this problem be solved?

Requiring algos and other trading systems to be designed to incorporate more than one source of reference price data is one possible solution. Another is to designate an alternative venue (either another primary market or an MTF) as the back-up reference price in the event of an outage. No doubt there are other solutions too. Both potential solutions described, require non-trivial technology work for market participants – which is perhaps why this problem exists. Market participants choose not to invest in the technology required to handle infrequent market outages. There is clearly a need for open transparent dialogue among market operators and participants about the scope of the problem, its costs, and the best possible solutions. In the absence of a market-led solution regulatory intervention may be necessary.

[1]. We exclude the closing auction from our analysis but note that the closing auction did not occur on 19 October due to another trading system issue.\ [2]. We report trading activity in logs to make it easy to identify differences in activity across the four venues.
[3]. and * denote significance at the 5%, and 1% level

Appendix

This appendix provides supporting materials to support our analysis. It is organised as follows:

  1. Variable definitions
  2. Details of the difference-in-difference approach
  3. Sample stocks
  4. Measuring price discovery

A1. Variable definitions

Our variables are measured in one minute increments between 9:00 and 17:30 CET. Variables are measured separately for the primary market, BATS, Chi-X and Turquoise.

Trading activity is trading volume measured in euros.

Bid ask spreads are the best ask price minus the best bid price divide by the midpoint of the best bid and ask price multiplied by 10,000 to convert to basis points.

Volatility is the standard deviation of one minute midpoint log returns measured in percentage.

Percentage of time at the best bid and offer without the primary market is calculated by observing the best prices on each market each minute. If a given MTF is at either the best bid and/or ask price and the primary market is not, this is counted as the MTF being at the best price. We sum up the number of observations and divide by 510 (the number of minutes in the trading day).

Price discovery is measured by the Hasbrouck information share (see Hasbrouck, 1995).

A2. Details of the difference-in-difference approach

The difference-in-difference regression model for our analysis, specified in levels is as follows:

\begin{align} y = \beta_0 + \beta_1 (Treatment \times Halt) + \beta_2 Treatment + \beta_3 Halt + \epsilon \end{align}

where $y$ is the bid ask spread measured in basis points, $Treatment$ is the treatment variable equal to one if the stock is listed on Euronext, and zero if listed elsewhere in Europe, $Halt$ is the halt indicator equal to one when the Euronext is halted and zero otherwise. Including the level $Treatment$ controls for differences between the treatment and control groups. Including the level $Halt$ controls for trends common to both treatment and control groups during the halting period. The variation that remains is the change in spreads experienced by Euronext stocks relative to the change in spreads for stocks listed in other European market. This variation is captured by $\beta_1$ the difference-in-difference estimate.

A3. Measuring price discovery

Hasbrouck (1995) developed a method to determine where price information or price discovery occurs when trading occurs in multiple venues. The approach assumes that there is a common implicit efficient price for a given security across all venues, and sources of variation in the efficient price are attributed to different venues. A venue’s contribution to price discovery is its information share, defined as the proportion of the efficient price innovation variance that can be attributed to that venue.

Specifically, if $p_{1,t}, p_{2,t}$ are prices from two venues (say the midquote from the primary market and the average midquote from alternative MTFs) tracking the same asset, then based on the law of one price, $p_{1,t}, p_{2,t}$ should be conintegrated (i.e., the linear combination of $p_{1,t}, p_{2,t}$ is a stationary process). We could capture the joint price dynamics with a vector error correction model (VECM), \begin{align} \Delta p_t = B_0 + B_1\Delta p_{t-1} + B_2\Delta p_{t-2} + ... + B_k\Delta p_{t-k} + \underbrace{\alpha \beta^T p_t}_{\text{the error correction term}} + \epsilon_t, \end{align} where $p_t$ is a vector of prices, i.e., $p_t = \left[\begin{array}{c}p_{1,t} \\ p_{2,t}\end{array}\right]$. In the above equation, $\alpha \beta^T p_t$ is the error correction term capturing the long-run (or equilibrium) dynamics between $p_{1,t}, p_{2,t}$.

Hasbrouck (1995) transforms the VECM model into a vector moving average (VMA) model, \begin{align} \Delta p_t = \sum_{\tau=0}^\infty\Psi_{\tau}\epsilon_{t-\tau}. \end{align} The VMA representation gives the impulse response function (IRF) subsequent to an arbitrary initial shock (from $\epsilon$). And $\sum_{\tau=0}^\infty\Psi_{\tau}$ captures the cumulative long-run predicted price changes implied by an initial shock.

Let $[\sum_{\tau=0}^\infty\Psi_{\tau}]_*$ denote any row of $\sum_{\tau=0}^\infty\Psi_{\tau}$, the Hasbrouck information share is defined as follows, \begin{align} IS_1 = \frac{d_1^2}{d_1^2 + d_2^2}, \\ IS_2 = \frac{d_2^2}{d_1^2 + d_2^2}, \end{align} where the vector of $d_i = [\sum_{\tau=0}^\infty\Psi_{\tau}]_* L, \text{ for }i = 1,2$, and $L$ is the Cholesky factor of the variance covariance matrix of the VECM, i.e., $\Sigma_\epsilon = LH$.

A4. Sample stocks

We examine stocks included in the STOXX Total Market Index (TMI) that trade on the primary market and the three largest Multilateral Trading Facilities (MTF): Turquoise, Chi-X and BATS, each day of our sample. Our sample includes 567 stocks: 109 of which have their primary listing on a Euronext market and 458 listed on non-Euronext markets. We refer to the non-Euronext sample as the control group. The following table contains the list of our sample stocks.